Classification of fall risk in Parkinson's disease using empirical mode decomposition and machine learning.
This study used empirical mode decomposition of center-of-pressure signals and machine learning (SFFS feature selection plus classifiers) to classify fall risk in 32 PD patients, identifying a 3-feature subset and achieving up to 0.96 subject-level AUC in the best condition.
What the AI sees
This study used empirical mode decomposition of center-of-pressure signals and machine learning (SFFS feature selection plus classifiers) to classify fall risk in 32 PD patients, identifying a 3-feature subset and achieving up to 0.96 subject-level AUC in the best condition.
Research significance
Offers a robust, objective monitoring/stratification tool for fall risk and potential clinical-trial endpoints in PD, improving assessment and remote monitoring, but provides little direct mechanistic or therapeutic discovery insight.
Source abstract
INTRODUCTION: Fall risk assessment in Parkinson's Disease (PD) is critical for preventing patient injuries; however, traditional clinical scales are often constrained by their semi-quantitative nature and time-consuming administration. This study aimed to develop and validate an analytical framework integrating Empirical Mode Decomposition (EMD) and machine learning for the objective classification of fall risk in PD patients. METHODS: Data from 32 PD patients under four standing conditions were analyzed, with fall risk labels defined based on a Mini-BESTest score cutoff of ≤21. Center of Pressure (COP) signals were first decomposed into Intrinsic Mode Functions (IMFs) using EMD, from which multi-dimensional features were extracted. Subsequently, the Sequential Forward Floating Selection (SFFS) algorithm was employed to identify a core feature subset, and the performance of five classifiers-Random Forest (RF), Decision Tree (DT), Extreme Gradient Boosting (XGB), Logistic Regression (LR), and Support Vector Machine (SVM)-was evaluated using Leave-One-Subject-Out Cross-Validation (LOSO-CV). Finally, a multi-trial probability aggregation strategy was introduced to assess subject-level risk. RESULTS: Results indicated that the rigid surface eyes-closed (rs ec) condition yielded the optimal discriminatory capability. The SFFS algorithm identified a parsimonious 3-dimensional feature subset (COPap_IMF1_BinEntropy, COPml_IMF1_STD, COPap_IMF2_TotalPower), which enabled tree-based classifiers to exhibit superior performance. Notably, DT and RF achieved highly balanced sensitivity and specificity, reaching a maximum subject-level AUC of 0.96. DISCUSSION: In conclusion, EMD-derived multi-scale features can precisely capture postural control deficits in PD, offering a promising technical solution for the future development of efficient and objective PD clinical monitoring tools.